AI Glossary
The complete dictionary of Artificial Intelligence
Multi-scale detection head
A component of object detectors that applies predictions at multiple levels of the feature pyramid, using specific layers for each scale to optimize the detection of objects of different sizes.
Adaptive Anchor
A technique where anchor boxes are generated dynamically based on the statistics of object dimensions in the dataset, thus improving the alignment between anchors and small objects.
Focal Loss
A modified loss function that reduces the contribution of easy, well-classified examples and focuses on difficult examples, particularly effective for handling the class imbalance inherent in small object detection.
Path Aggregation Network - PANet
An architecture that improves information flow in feature pyramid networks by creating shortcuts between lower and upper layers, better preserving localization information for small objects.
Detection-guided Super-resolution
An approach that increases the resolution of feature maps or regions of interest before classification, using super-resolution techniques to make small objects more discernible.
Feature Dilatation
A technique that increases the receptive field of neurons without losing spatial resolution, allowing to capture more context around small objects while maintaining their precise localization.
Keypoint Sampling
A method that selectively samples features from the most informative points of a small object region, rather than using global pooling that might drown important details.
Spatial Attention Module
A mechanism that learns to weight the importance of different spatial regions of the feature map, allowing the network to focus on areas containing relevant small objects.
Feature Re-identification
A process that re-injects low-level features into deeper layers of the network, helping to recover fine details lost during successive downsampling operations.
Multi-scale Context
The use of contextual information from different spatial scales around a small object to improve its classification, by exploiting the relationships between the object and its environment.
Small object-adapted data augmentation
Image transformation techniques specifically designed to increase the presence and variability of small objects in the training dataset, such as selective zooming or copy-pasting objects.
Size-weighted IoU
An evaluation metric that gives more weight to correct predictions on small objects when calculating the Intersection over Union, better reflecting the difficulty and importance of their detection.
Cascade Detection Head
An architecture where several detection heads are placed in sequence, each head refining the predictions of the previous one with increasingly strict IoU thresholds, improving accuracy on difficult objects such as small objects.
Bi-directional feature fusion
A technique that combines high-level (semantic) and low-level (spatial) features in both directions, so that deep layers receive precise localization information to better detect small objects.
Feature Highlighting
A method that amplifies the responses of neurons activated by small patterns or textures, making the features of small objects more salient compared to background noise.
Contrastive learning for detection
A training approach that teaches the network to bring the features of similar small objects closer together and push those of different objects or the background further apart, improving their separability.